Thinky 发布 Inkling：975B 总参数 / 41B 激活参数多模态模型，美国最强 Apache 2.0 开源模型（附 Inkling-Small，276B/12B）

Thinky 似乎每隔几个月才浮出水面一次；最近一次是关于交互模型 —— 但每次亮相都令人印象深刻，既有品味又有深度。今天他们推出了 Inkling —— 虽非 SOTA 模型，但却是为美国开源模型基线打造的一个非常扎实的新系列：

• 我们的模型名为 Inkling，是一个混合专家 Transformer，总参数 975B，激活参数 41B。
• 支持高达 1M token 的上下文窗口。
• 使用 45 万亿 token 的文本、图像、音频和视频数据进行预训练。
• 这是不同尺寸模型家族的首款作品：同时我们分享 Inkling-Small 预览，一个仅 12B 激活参数的轻量模型，采用相似配方训练，以更低的成本和延迟实现强劲性能。
• Inkling 原生对文本、图像和音频进行推理，并通过高效且可控的思考努力来平衡成本与性能。

Huggingface 拆解 介绍了一些有趣的技术亮点：

Thinking Machines Lab 推出了 Inkling，它是其首个完全发布的开源权重基础模型家族的首款作品，定位为可定制的多模态基座模型，而非追求最高基准的旗舰产品。

• Thinking Machines 宣布 Inkling 为一个开源权重模型，可"跨文本、图像和音频模态高效推理"，提供完整权重，并在其 Tinker 平台和 Playground 上立即提供支持 @thinkymachines。

• Mira Murati 将 Inkling 描述为公司的"第一个模型"，"从零开始训练"，拥有开源权重，且同一天即可在 Tinker 上微调 @miramurati。

• Soumith Chintala 将其定位为 Thinking Machines 的"首个通用模型"，强调开源权重、975B 参数、原生多模态，以及在 Tinker、Hugging Face 和合作伙伴平台上的可用性 @soumithchintala。

• John Schulman 补充了时间线：预训练始于去年冬天，并从一月中旬开始，一个小团队在此基础上构建了编码、推理和智能体训练 @johnschulman2。

• Lilian Weng 将 Inkling 描述为旨在"在广泛能力类别上实现稳健性能"的基础模型，适用于实际使用和定制 @lilianweng。

• TML 团队成员反复强调，这是首发版本，为未来迭代奠定基础，而非最终极限突破 @soumithchintala, @cHHillee, @keirp1。

• 此次发布得到了异常广泛的首日生态支持，涵盖 vLLM、SGLang、Modal、Baseten、Databricks、Hugging Face 以及量化/社区工具 @vllm_project, @lmsysorg, @modal, @baseten, @Yuchenj_UW, @huggingface, @danielhanchen。

• 独立评论者立即将其标记为迄今为止最强的美国开源权重版本，尽管在某些基准上通常仍落后于顶级中国开源权重和最佳闭源模型 @natolambert, @ArtificialAnlys, @scaling01。

• 多数帖子报道 Inkling 为 975B 总参数 / 41B 激活参数 @soumithchintala, @vllm_project, @ArtificialAnlys, @kimmonismus。

• 它是一个 混合专家 模型，每个 token 激活 41B 参数 @VictoriaLinML。

• 根据多方反应和总结，采用 Apache 2.0 许可 @natolambert, @Yuchenj_UW, @multimodalart。

• 支持 文本、图像和音频输入，输出 文本 @soumithchintala, @TheRundownAI, @ArtificialAnlys。

• 开源权重检查点支持高达 1M 上下文 @vllm_project, @lmsysorg, @ArtificialAnlys。

• Tinker/API 上下文被描述为 256K，针对 64K 和 256K 上下文有不同的定价 @ArtificialAnlys。

若干具有技术洞察力的反应从发布中提取了架构选择：

• 混合/滑动窗口注意力，采用 5:1 局部与全局层比例，窗口大小为 512 @eliebakouch, @ariG23498。

• 相对位置编码 / 相对注意力偏置 代替 RoPE；多位发帖者称这是最大胆的新颖选择之一 @stochasticchasm, @eliebakouch, @rasbt, @arohan , @ChangJonathanC。

• 短卷积层 被添加在注意力/FFN 流周围；评论者指出这是短卷积异常大规模的应用 @eliebakouch, @stochasticchasm, @rasbt, @SonglinYang4。

• 具有共享专家接收器的 MoE / 2 个共享专家，社区认为这不常见，因为近年来许多 MoE 仅使用 1 个共享专家 @eliebakouch, @ariG23498。

• 社区对架构的解读中引用了 DeepSeek 风格的无辅助损失负载均衡 @eliebakouch。

• 从技术撰写中推断出使用了 muP 以及 Muon/权重衰减变体，并得到优化器专家的确认：Aaron Defazio 表示他们正使用其修正的权重衰减方法"MuonC/AdamC" @aaron_defazio，同时社区读者也指出了 muP @stochasticchasm, @Laz4rz。

• vLLM 强调了用于推测解码的 8 个 MTP 头 @vllm_project。

• Inkling-Small 被反复提及为即将到来或单独讨论的较小模型 @LiorOnAI, @teortaxesTex。

• 社区摘要描述 Inkling-Small 为 276B 总参数 / 12B 激活参数，并在多项评估中出人意料地与更大模型竞争 @eliebakouch, @nrehiew_。

• Artificial Analysis 表示 Inkling 以 41 分的智能指数 首次亮相，成为领先的美国开源权重版本，超过 Nemotron 3 Ultra (38)、Gemma 4 31B (29) 和 gpt-oss-120b (24) @ArtificialAnlys。

• Artificial Analysis 还指出，Inkling 每个智能指数任务平均输出 25K token，对比 GLM-5.2 max 的 43K、Kimi K2.6 的 38K 和 DeepSeek v4 Pro max 的 37K，显示其相对节省 token @ArtificialAnlys。

• Natolambert 称其为"相比 Nemotron Ultra 有明显的进步"和"新的最佳美国模型"，但"在智能体基准上仍稍落后于 GLM 5.2，在多模态上落后于 Kimi K 2.6" @natolambert。

• Design Arena 指出 Inkling 进入智能体 Web 应用竞技场 总排名第 9，Elo 1257，与 Claude Opus 4.6 和 Gemini 3.5 Flash 同档，并称其为针对智能体工作负载排名最高的美国开源权重模型 @DesignArena。

• 竞技场在发布日将 Inkling 添加至 Agent Arena / Text / Vision / Code Arena @arena。

来自 Artificial Analysis：

• GDPval-AA v2 Elo 1238，高于 Kimi K2.6 (1190) 和 DeepSeek v4 Flash max (1189) @ArtificialAnlys。
• τ³-Banking 24%，高于 Kimi K2.6 (21%)，略高于 DeepSeek v4 Flash max (23%) @ArtificialAnlys。

正面：

• "犀利而简洁"的推理，不啰嗦 @MichaelElabd。
• 工具调用能力强，在智能体任务上具有良好的长周期错误恢复能力 @MichaelElabd。
• 良好的"思维品质" / 不阿谀的风格 @skirano, @tinkerapi。
• Alex Kirillov 声称 Inkling 避免了许多全模态模型中常见的"音频输入导致智能惩罚"，但另一位用户要求更有力的证据和基准 @alex_kirillov , @giffmana, @alex_kirillov。

更偏混合/批评：

• Scaling01 认为基准测试"并不那么出色"，描述其大致为"另一款 Kimi-K2.6"，落后所有闭源模型和 GLM-5.2，并推测此次发布可能是赶在 Kimi-K3 和 DeepSeek-V4-GA 之前 @scaling01。

• Stochasticchasm 表示它"在多模态上似乎非常强"但"在终端基准等上并非特别强" @stochasticchasm。

• JJitsev 反驳了关于"唯一未使用蒸馏训练的开源权重模型"的炒作，称 Inkling 使用了开源权重的蒸馏，并在 TerminalBench 风格评估中表现不如 GLM 5.2 @JJitsev。

• TeortaxesTex 提供了一个逆向的积极观点：平庸的基准刷分可能表明更少的捷径/蒸馏污染和更独立的数据管线 @teortaxesTex。

• NVIDIA 表示 Inkling 是在 GB300 NVL72 上训练的，并且 NVFP4 检查点 在发布当天即已在 Hugging Face 上提供 @NVIDIAAI。

• vLLM 称首日支持包括 NVFP4 和 BF16，针对 Blackwell 和 Hopper 优化，在 4× GB200 上使用 MTP 可达到 380 tok/s/user @vllm_project。

• Inferact 详述了系统工作：能够感知短卷积的张量并行分片、低延迟融合集合通信（bs=1 时快 5 倍），以及直接集成 TML 的 FA4 剪切偏置核 @inferact。

• LMSYS/SGLang 表示原生实现了 Inkling 架构支持，包括 ShortConv、相对位置注意力、共享专家接收器 MoE、预填充全 CUDA 图、MXFP8 KV 缓存、在定制 Megatron 后端中的全参数和 LoRA RL、路由回放、跨运行时参数同步，以及来自 Modal 的 DFlash 推测解码 @lmsysorg。

• Modal 称，Inkling 在 Modal 上使用了一个 c

Thinky only seems to come up for air once every few months; most recently with Interaction models - but each time they do they impress, showing both taste and depth. Today they introduced Inkling — not a SOTA model, but a very solid new family for a baseline American open model:

• Our model, called Inkling, is a Mixture-of-Experts transformer with 975B total parameters, 41B active.

• It supports a context window of up to 1M tokens.

• It was pretrained on 45 trillion tokens of text, images, audio and video.

• It is the first in a family of models of different sizes: alongside it we are sharing a preview of Inkling-Small, a lighter-weight model with 12B active parameters, trained with a similar recipe, that achieves strong performance with even lower cost and latency.

• Inkling reasons natively over text, images, and audio, and balances cost with performance through efficient and controllable thinking effort

The Huggingface breakdown covers some interesting technical highlights:

AI News for 7/14/2026-7/15/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

Thinking Machines Lab launched Inkling, its first fully released open-weights foundation model family entry, positioning it as a customizable multimodal base model rather than a benchmark-maxed flagship.

• Thinking Machines announced Inkling as an open-weights model that “reasons efficiently across text, image, and audio modalities,” with full weights available and immediate support on its Tinker platform and Playground @thinkymachines.

• Mira Murati described Inkling as the company’s “first model,” “trained from scratch,” with open weights and same-day fine-tuning on Tinker @miramurati.

• Soumith Chintala framed it as Thinking Machines’ “first general model,” stressing open weights, 975B parameters, native multimodality, and availability on Tinker, Hugging Face, and partners @soumithchintala.

• John Schulman added timeline context: pretraining began last winter, and from mid-January a small team built coding, reasoning, and agentic training on top @johnschulman2.

• Lilian Weng characterized Inkling as a foundation model aimed at “solid performance across a broad categories of capabilities” and intended for practical use plus customization @lilianweng.

• TML staff repeatedly emphasized that this is a day-1 release and a foundation for future iterations rather than their final frontier push @soumithchintala, @cHHillee, @keirp1.

• The release landed with unusually broad day-0 ecosystem support across vLLM, SGLang, Modal, Baseten, Databricks, Hugging Face, and quantization/community tooling @vllm_project, @lmsysorg, @modal, @baseten, @Yuchenj_UW, @huggingface, @danielhanchen.

• Independent commentators immediately tagged it as the strongest U.S.-based open-weight release so far, though generally still behind the top Chinese open-weight and best closed models on some benchmarks @natolambert, @ArtificialAnlys, @scaling01.

• Inkling is reported as 975B total parameters / 41B active parameters in most posts @soumithchintala, @vllm_project, @ArtificialAnlys, @kimmonismus.

• It is a Mixture-of-Experts model with 41B active parameters per token @VictoriaLinML.

• It is Apache 2.0 licensed according to multiple reactions and summaries @natolambert, @Yuchenj_UW, @multimodalart.

• It supports text, image, and audio inputs , with text output @soumithchintala, @TheRundownAI, @ArtificialAnlys.

• Open-weights checkpoints support up to 1M context @vllm_project, @lmsysorg, @ArtificialAnlys.

• Tinker/API context is described as 256K , with pricing differentiated for 64K and 256K contexts @ArtificialAnlys.

Several technically literate reactions extracted architectural choices from the release:

• Hybrid/sliding-window attention with a 5:1 local-to-global layer ratio and window size 512 @eliebakouch, @ariG23498.

• Relative positional encoding / relative attention bias instead of RoPE; multiple posters called this one of the most novel large-scale choices @stochasticchasm, @eliebakouch, @rasbt, @arohan , @ChangJonathanC.

• Short convolution layers added around attention/FFN streams; commenters flagged this as unusually scaled-up usage of short convs @eliebakouch, @stochasticchasm, @rasbt, @SonglinYang4.

• MoE with shared expert sinks / 2 shared experts , noted as atypical since many recent MoEs use 1 shared expert @eliebakouch, @ariG23498.

• DeepSeek-style auxiliary-loss-free load balancing was cited in community readings of the architecture @eliebakouch.

• muP and Muon/weight decay variants were inferred from the writeup and confirmed by optimizer expert reaction: Aaron Defazio said they are using his corrected weight decay approach, “MuonC/AdamC” @aaron_defazio, while community readers also pointed out muP @stochasticchasm, @Laz4rz.

• 8 MTP heads for speculative decoding were highlighted by vLLM @vllm_project.

• Inkling-Small is repeatedly referenced as an upcoming or separately discussed smaller model @LiorOnAI, @teortaxesTex.

• Community summaries describe Inkling-Small as 276B total / 12B active and unexpectedly competitive versus the larger model on several evaluations @eliebakouch, @nrehiew_.

• Artificial Analysis said Inkling debuts at 41 on the Intelligence Index , making it the leading U.S. open-weights release and ahead of Nemotron 3 Ultra (38) , Gemma 4 31B (29) , and gpt-oss-120b (24) @ArtificialAnlys.

• Artificial Analysis also said Inkling averages 25K output tokens per Intelligence Index task , vs 43K for GLM-5.2 max , 38K for Kimi K2.6 , and 37K for DeepSeek v4 Pro max , framing it as relatively token-efficient @ArtificialAnlys.

• Natolambert called it a “clear step up from Nemotron Ultra” and “new best American model,” but still “a bit behind GLM 5.2 on agentic benchies, and Kimi K 2.6 on multi modal” @natolambert.

• Design Arena said Inkling entered Agentic Web App Arena at #9 overall, Elo 1257 , in the same band as Claude Opus 4.6 and Gemini 3.5 Flash , and called it the highest-ranking U.S.-based open-weight model for agentic workloads @DesignArena.

• Arena added Inkling to Agent Arena / Text / Vision / Code Arena on launch day @arena.

From Artificial Analysis:

• GDPval-AA v2 Elo 1238 , higher than Kimi K2.6 (1190) and DeepSeek v4 Flash max (1189) @ArtificialAnlys.

• τ³-Banking 24% , above Kimi K2.6 (21%) and slightly above DeepSeek v4 Flash max (23%) @ArtificialAnlys.

Positive:

• “Sharp and concise” reasoning, not rambly @MichaelElabd.

• Strong tool calling and good long-horizon error recovery on agentic tasks @MichaelElabd.

• Good “quality of mind” / unsycophantic flavor @skirano, @tinkerapi.

• Alex Kirillov claimed Inkling avoids the common “audio in = intelligence penalty” seen in many omni models, though another user asked for stronger supporting evidence and benchmarks @alex_kirillov , @giffmana, @alex_kirillov.

More mixed / critical:

• Scaling01 argued the benchmarks are “not that great,” describing it as roughly “another Kimi-K2.6” and behind all closed models and GLM-5.2, speculating the release may have been timed ahead of Kimi-K3 and DeepSeek-V4-GA @scaling01.

• Stochasticchasm said it seems “very strong for multimodal” but “not super strong for terminal bench etc.” @stochasticchasm.

• JJitsev pushed back on hype around “only open-weight model trained without distilling,” saying Inkling uses distillation from open weights and underperforms GLM 5.2 on TerminalBench-style evals @JJitsev.

• TeortaxesTex offered a contrarian positive spin: mediocre benchmark-maxing may actually suggest less corner-cutting/distillation contamination and a more independent data pipeline @teortaxesTex.

• NVIDIA said Inkling was trained on GB300 NVL72 and that an NVFP4 checkpoint was available on Hugging Face on day 0 @NVIDIAAI.

• vLLM said day-0 support includes NVFP4 and BF16 , optimized for Blackwell and Hopper , reaching up to 380 tok/s/user on 4× GB200 with MTP @vllm_project.

• Inferact detailed system work: sconv-aware tensor-parallel sharding , low-latency fused collectives (5× faster at bs=1) , and direct integration of TML’s FA4 sheared-bias kernel @inferact.

• LMSYS/SGLang said Inkling architecture support was implemented natively, including ShortConv , relative positional attention , shared expert sink MoE , prefill full CUDA graph , MXFP8 KV cache , full parameter and LoRA RL in customized Megatron backend , routing replay , cross-runtime parameter sync , and DFlash speculative decoding from Modal @lmsysorg.

• Modal said Inkling on Modal uses a custom DFlash speculator for 67% higher throughput and interactivity @modal.

• Soumith Chintala separately amplified that Modal’s DFlash speculator is “much faster than MTP” @soumithchintala.

• Lysandre reported replacing TML’s causal Conv1D with causal-conv1d yielded +4% tok/s , and replacing attention with FlashAttention-4 yielded another +11% , for ~15% total throughput gain without retraining @LysandreJik.

• Unsloth released 1-bit GGUF quants said to be 86% smaller (270GB vs 1.9TB) while retaining 74.2% of top-1% accuracy , with vision and audio support @danielhanchen.

• Artificial Analysis listed Tinker pricing as:

  • 64K context : $1.87 / 1M input , $0.374 cached , $4.68 output

  • 256K context : $3.74 / 1M input , $0.748 cached , $9.36 output
    @ArtificialAnlys

• Available on Tinker , Hugging Face , and via launch partners including Databricks , Baseten , Modal , vLLM/SGLang stacks @soumithchintala, @Yuchenj_UW, @baseten, @modal.

• “Best American open model” / “saved American open-source frontier” are judgments, albeit repeated by several respected observers @natolambert, @karinanguyen, @saranormous.

• Claims that Inkling is especially important because it is not distilled from OpenAI/Anthropic are disputed. Jxmnop called it “the ONLY open-weight model” without such distillation @jxmnop, then partially walked it back: “apparently they did distill lol. but only a tiny bit” @jxmnop. Andrew Carr also contested the purity framing, noting use of Kimi 2.5 for SFT traces @andrew_n_carr.

• Claims that Inkling was “rushed” ahead of Chinese releases are speculation from critics, not evidenced by the launch materials @scaling01.

• Claims that relative attention gives TML a finetuning moat because backward is hard are speculative @typedfemale.

• Claims that Inkling avoids multimodal intelligence loss are promising but not yet benchmark-complete in the tweet set @alex_kirillov.

• Open-weight and permissive license as strategic win: Many saw the Apache-2.0 release as a major boost to the U.S./Western open ecosystem @latkins, @saranormous, @brexton, @hyperindexed.

• Customization over leaderboard chasing: Researchers and builders praised the explicit framing that Inkling is a broad, tunable foundation rather than a benchmark-maxed point solution @gneubig, @ben_burtenshaw, @thealexker.

• Strong release quality: Several users praised the transparency, grounded tone, and comprehensive technical documentation @lvwerra, @saranormous, @rasbt.

• Architecture interest: The non-RoPE positional choice and scaled short-conv usage drew positive attention as evidence TML is willing to make meaningful architecture bets @stochasticchasm, @rasbt, @ChangJonathanC.

• Strong but not top overall: The most balanced reads place Inkling as the new U.S. open-weight leader, but behind GLM/Kimi/DeepSeek or top closed models on some fronts @natolambert, @ArtificialAnlys, @stochasticchasm.

• Good base model thesis: Multiple analysts read the release as a systems/business move: ship a solid, efficient, post-trainable base and let Tinker plus downstream RL/fine-tuning create differentiation @ben_burtenshaw, @kimmonismus, @tinkerapi.

• Not frontier overall: Critics argued it is still clearly behind top Chinese open-weight models and the strongest closed models @scaling01, @JJitsev.

• Purity claims overstated: Some pushback focused on exaggerated claims that it is uniquely “pure” or non-distilled; the thread set includes both hype and corrections @jxmnop, @jxmnop, @andrew_n_carr, @JJitsev.

• Benchmark middlingness as concern: Some readers saw the moderate benchmark profile as evidence it may simply lag current Chinese open frontier rather than inaugurate a new frontier @scaling01.

• First major TML public model: This is the first true external model release from Thinking Machines after months of anticipation around a lab staffed by ex-OpenAI leaders and researchers. That made the choice of open weights itself notable @Hesamation, @TechCrunch.

• A U.S. open-weight answer to Chinese momentum: Many reactions explicitly compare Inkling to GLM, Kimi, DeepSeek, and Qwen. The release lands amid concern that Western open-weight models have trailed Chinese ones on capability and release cadence @scaling01, @teortaxesTex, @sriramk.

• Open base + post-training stack thesis: TML’s messaging strongly suggests a strategy similar to “ship a competent open substrate, then differentiate via customization/fine-tuning/RL infrastructure.” That aligns with Tinker distribution and with user reactions centering controllable reasoning, concise outputs, and adaptation rather than raw leaderboard supremacy @thinkymachines, @MichaelElabd, @ben_burtenshaw.

• Inference ecosystem maturity: The release also showcases how far open inference stacks have come. Day-0 support for a 1T-class multimodal MoE with new architectural components and multiple kernel-level optimizations would have been far less plausible a year earlier @vllm_project, @inferact, @LysandreJik.

• Architectural experimentation at scale: Relative positional bias instead of RoPE and large-scale short-conv usage are the kind of choices researchers watch closely because they may indicate future architecture trends if they prove robust under scaling and post-training @stochasticchasm, @rasbt, @ChangJonathanC.

• Release style as signal: Several commentators praised the unusually restrained release language, explicit admission that it is not the strongest overall model, and detailed technical notes. For expert audiences, that improved credibility relative to more benchmark-maxed launches @eliebakouch, @lvwerra, @thealexker.